Підходи до пояснюваності рішень великих мовних моделей у інтелектуальних системах

Abstract

This work is devoted to the analysis of explainability approaches in Large Language Models used within intelligent information systems. The growing complexity of transformer-based architectures strengthens the “black-box” problem and increases the need for transparent and interpretable decision-making. The study examines post-hoc interpretation methods, attention-based analysis, and chain-ofthought prompting as practical tools for improving model transparency. Special attention is paid to emerging directions of mechanistic interpretability aimed at understanding internal representations of LLMs. The results emphasize the necessity of combining different interpretability strategies to enhance reliability, auditability, and user trust in LLM-based applications